An AI draft can be clear and still sound wrong for your brand. Give the writer a short reference with approved examples, then review the draft against it before publishing.
The Brand Voice OS below is a five-part editorial framework for doing that. It gives your team something concrete to discuss and revise; it does not guarantee that a model will follow every instruction.
What goes into a Brand Voice OS?
Keep the reference easy to find, dated and owned by someone who can approve changes. Start with these five parts and expand only when a real drafting problem calls for more detail.
- Audience and situation. Say who is reading, what they need and how they may feel. A product announcement and an apology can share a voice while using different tones.
- Checkable writing rules. Turn a broad principle into an action. For example, “helpful” can mean “state the next step and explain any unfamiliar term.” Do not force a number into every claim merely to sound confident.
- Preferred wording and boundaries. Record product names, useful terms and phrases to avoid. Keep approved factual claims alongside the voice notes, with their sources and dates.
- Approved examples with reasons. Pair a weak draft with a reviewed version of the same message. Explain what changed. Include different situations rather than several examples of the same kind of post.
- A review owner and change log. Name who checks accuracy, tone and destination before release. Save accepted edits so the next draft starts with better examples.
Mailchimp’s public guide illustrates the distinction between a relatively stable voice and tone that changes with context. Microsoft’s guide connects its principles to practical writing choices. Use those as examples of specificity, then define your own voice.
How specific should tone guidance be?
Write an anchor for each important choice. “Conversational” could mean contractions and direct address, with no slang. “Calm” could mean explaining the next step before using reassurance. A score such as 7/10 needs examples to explain what the score means.
Nielsen Norman Group’s four tone dimensions are humor, formality, respectfulness and enthusiasm. They offer a vocabulary for discussion. They are not a machine-checked standard or proof that a draft will suit your audience.
A worked example: explain a delivery delay
Fictional example. An online stationery shop needs to tell a customer that their notebook will arrive late. Its approved facts are: dispatch is delayed by two days, the customer will receive a tracking email after dispatch, and support can discuss cancellation. These facts are invented solely to demonstrate the method.
Draft that needs review
Exciting news! Your premium notebook is almost on its way. We guarantee it will be worth the wait. Stay tuned for an amazing delivery experience!
Reviewed version
Your notebook will dispatch two days later than planned. We’re sorry for the delay. We’ll email your tracking link after dispatch. If the new timing doesn’t work for you, contact support to discuss cancellation.
Why the edit works for this brief: it leads with the delay, uses the approved timing, removes an unsupported guarantee and gives the customer a next step. The shop’s “warm” voice does not require enthusiasm when a customer is inconvenienced.
Save both versions with that explanation. A reviewer can now point to specific choices instead of asking the next writer to “sound more human.” This is an editorial example, not a measured customer result.
What should the AI drafting prompt include?
Supply the current reference with the actual task, approved facts and relevant examples. Follow your organisation’s data rules before sending customer or confidential material to an AI service. Provider features and instruction handling differ; check the tool’s current documentation.
Task: Write [format] for [audience and situation]. Reader's next step: [one useful action]. Approved facts and sources: [dated facts; no new promises]. Voice reference: [current rules, preferred wording and boundaries]. Examples: [approved messages with reasons for their wording]. Constraints: [channel, length, required disclosures]. If a fact is missing, identify it for the reviewer rather than inventing it. Return a draft and a short list of claims the reviewer must check.
Bracketed fields above are template inputs to fill before use. Anthropic’s prompting guidance discusses clear instructions and relevant examples. The prompt here is our editorial template, not a provider-certified formula.
When starting a new task, check that the current voice reference is actually available to the tool. Saved instructions, projects and uploaded files behave differently across products. Revisit the reference when the task, audience or model changes. There is no fixed draft number at which voice drift begins.
How do you test the draft before publishing?
Keep a small set of representative tasks, such as an announcement, a delay message and a product explanation. Compare drafts with approved examples of similar work. A reviewer should record the actual reason for each edit.
- Accuracy: can each factual claim be traced to an approved source? Check names, dates, prices and promises separately from style.
- Voice: does the wording follow the written rules and fit the reader’s situation? Point to the line that needs changing.
- Usefulness: does the reader understand the message and next step?
- Release: is the format, link and destination correct, and has the responsible person approved it?
Record recurring edits by category. If drafts repeatedly add hype, improve the example and the instruction, then try the same test tasks again. If the reference is missing or instructions conflict, fix that first. Do not treat a model’s own “on-brand” verdict as independent validation.
For tools, compare the features you need, such as shared references, permissions and review workflow, rather than assuming every product is equivalent. Read our guide to brand voice AI tools. Customer-facing chatbots need separate factual, safety and escalation checks; see the chatbot voice guide.
Frequently asked questions
How long should the reference be? Start with a short usable document. Add enough examples to cover the situations you need; there is no universal page count or example threshold.
Can existing brand guidelines help? Yes. Extract the relevant rules, approved claims and examples, then test whether a writer can apply them to a real task. A complete style guide can already contain useful voice guidance.
Does it work with every AI tool? The text can be adapted for different tools, but instruction handling, access controls and saved context vary. Test your actual tool and workflow before relying on the result.
How will we know it helps? Track the kinds of edits reviewers make on comparable tasks. Look for fewer recurring voice corrections while maintaining accuracy and usefulness. Do not infer business impact from a few approved drafts.
When should we update it? Review it when positioning, audience, product facts or recurring edit patterns change. Assign an owner and a review date rather than relying on memory.
Put the reference to work
Choose one message you already need to write. Fill the Brand Voice OS Worksheet in the resource library, draft the message and have a reviewer explain the edits. For a documented example of the wider workflow, open the worked marketing decisions.
Editorial method reviewed 29 October 2026. The fictional example demonstrates writing choices; no conversion or consistency improvement has been measured here.